From 078f7cd45196a01a47539ddc3a5bf106f86a00d5 Mon Sep 17 00:00:00 2001 From: cbb <2625274797@qq.com> Date: Tue, 20 Oct 2026 22:24:49 +0800 Subject: [PATCH] =?UTF-8?q?feat+test:=20select=3Dep=5Fwaste=20=E7=99=BD?= =?UTF-8?q?=E8=B4=B9=E6=AF=94=E4=BE=8B=E6=83=A9=E7=BD=9A(=E7=BD=9A?= =?UTF-8?q?=E5=BA=A6=E6=8C=89=E5=BD=93=E5=89=8D=E8=AF=AF=E5=B7=AE=E7=BC=A9?= =?UTF-8?q?=E6=94=BE)=EF=BC=9B=E5=8D=95=E6=B5=8B=E6=8A=93=E5=88=B0?= =?UTF-8?q?=E6=88=91=E6=8A=8A=E5=87=80=E4=BD=8D=E7=A7=BB=E5=AF=B9=E6=97=B6?= =?UTF-8?q?=E9=97=B4=E8=BD=B4=E6=B1=82=E5=92=8C=E5=86=99=E6=88=90=E4=BA=86?= =?UTF-8?q?=E5=AF=B9=E8=BD=B4=E6=B1=82=E5=92=8C(frac=20=E6=81=92=200)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- ml/eval/eval_cl.py | 2 +- ml/sim/closed_loop.py | 10 ++++++++-- tests/test_eval_cl.py | 17 +++++++++++++++-- 3 files changed, 24 insertions(+), 5 deletions(-) diff --git a/ml/eval/eval_cl.py b/ml/eval/eval_cl.py index bfcd66d..d48c722 100644 --- a/ml/eval/eval_cl.py +++ b/ml/eval/eval_cl.py @@ -98,7 +98,7 @@ def main(): ap.add_argument('--cert_batch', type=int, default=0, help='1=一次 gaps 调用覆盖全部候选（压最坏情况，典型路径略贵）') ap.add_argument('--steps', type=int, default=1, help='MeanFlow 采样段数；1=线上一口径（历史发布数字全是 1）') ap.add_argument('--z_scale', type=float, default=1.0, help='候选噪声缩放：>1 让 K 条候选更分散（测"全拒是没有备选"还是"真的没路"）') - ap.add_argument('--len_w', type=float, default=0.0, help="select=ep_len 时行程惩罚权重（mm / 每轴归一 STEP 的累计路程）") + ap.add_argument('--len_w', type=float, default=0.0, help="ep_len=罚累计归一行程；ep_waste=罚(1−净/总)×当前误差；0 与 endpoint 逐位相同") ap.add_argument('--backoff', type=int, default=0, help='全拒时把首选候选按 1/2、1/4…缩步重试的档数（需 --cert_batch 1）') ap.add_argument('--max_sec', type=float, default=None) ap.add_argument('--seed', type=int, default=0) diff --git a/ml/sim/closed_loop.py b/ml/sim/closed_loop.py index 4728fe8..f91d095 100644 --- a/ml/sim/closed_loop.py +++ b/ml/sim/closed_loop.py @@ -284,13 +284,19 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None, fut = qe[:, None, None] + torch.cumsum(inc[:, :, r:], 2) # [B,K,H-r,6] if select == 'first': score = torch.zeros(B, K, device=device); score[:, 0] = 1 - elif select in ('endpoint', 'adapt', 'ep_len'): + elif select in ('endpoint', 'adapt', 'ep_len', 'ep_waste'): qa = fut[:, :, -1].reshape(B * K, 6) e_end, _ = goal_err(kin, qa, T.repeat_interleave(K, 0), R.repeat_interleave(K, 0)) score = -e_end.view(B, K) - if select == 'ep_len' and len_w > 0: # 行程惩罚：每轴按 STEP 归一的累计路程。§9.2 要"行程≤规划器×1.10"， + if select == 'ep_len' and len_w > 0: # 行程惩罚：每轴按 STEP 归一的累计路程。§9.2 要“行程≤规划器×1.10”， ln = ((fut[:, :, 1:] - fut[:, :, :-1]).abs() / step).sum(-1).sum(-1) # 而现在 2.4 倍没人管 score = score - len_w * ln + if select == 'ep_waste' and len_w > 0: # 轮次 39 的分解证明行程超出主要是同一轴来回抖（换向步占 23.5%、 + dv = fut[:, :, 1:] - fut[:, :, :-1] # 净位移只占总行程 44.5%），绕远只占一小部分 ⇒ 罚"白费的比例"： + tot = dv.abs().sum(-2).sum(-1) # 同轴来回会在净位移里相互抵消，所以 waste = 1 − 净/总 ∈ [0,1]。 + net = dv.sum(-2).abs().sum(-1) # dv 是 [B,K,H-1,6]：时间在第 -2 轴（写错成 -1 就变成"对轴求和"，等于总行程） + frac = (1.0 - net / tot.clamp_min(1e-6)).clamp(0, 1) + score = score - len_w * frac * e.view(B, 1) # 罚度按"当前离目标多远"缩放：近端不会把进度项压死 if select == 'adapt' and last is not None: n = min(fut.shape[2], last.shape[1] - (commit_n + 1)) w = 0.85 ** torch.arange(n, device=device, dtype=torch.float32) diff --git a/tests/test_eval_cl.py b/tests/test_eval_cl.py index 629e2da..a1ac3ca 100644 --- a/tests/test_eval_cl.py +++ b/tests/test_eval_cl.py @@ -795,7 +795,7 @@ class TravelPenaltySelectTest(unittest.TestCase): return (self.act.expand(B, K, 16, 6).clone(), torch.zeros(B, device=feat.device), torch.zeros(B, 13, device=feat.device)) - def run1(self, len_w): + def run1(self, len_w, select='ep_len'): class SafeKin: def clamp_box(self, q): return q @@ -811,7 +811,7 @@ class TravelPenaltySelectTest(unittest.TestCase): patch('ml.sim.closed_loop.goal_err', side_effect=lambda kin, q, T, R: (torch.full((q.shape[0],), 10.0), torch.zeros(q.shape[0]))): _, per = rollout(self.StubModel(acts), SafeKin(), tasks, 'cpu', K=2, safety=True, max_sec=0.06, - return_traj=True, use_sub=False, select='ep_len', replan='empty', commit_n=2, + return_traj=True, use_sub=False, select=select, replan='empty', commit_n=2, safety_partial=1, len_w=len_w) return per @@ -821,6 +821,19 @@ class TravelPenaltySelectTest(unittest.TestCase): p1 = self.run1(1.0) # 加惩罚：行程 10.8 vs 3.6 ⇒ 选匀速那条 self.assertAlmostEqual(p1['traj'][0, 1, 0].item(), 0.225 * 1.2, places=5) + def test_waste_penalty_also_separates_them(self): + """换成"白费比例"（1 − 同轴净位移/累计行程）也必须能分开，且罚度按当前误差缩放： + 这里 goal_err 恒为 10 ⇒ 抖动条 waste≈0.78、匀速条 0 ⇒ 差 7.8×len_w 的当量。""" + p0 = self.run1(0.0, select='ep_waste') + self.assertAlmostEqual(p0['traj'][0, 1, 0].item(), 0.9 * 1.2, places=5) + p1 = self.run1(1.0, select='ep_waste') + self.assertAlmostEqual(p1['traj'][0, 1, 0].item(), 0.225 * 1.2, places=5) + + def test_zero_weight_ep_waste_is_the_old_endpoint_path(self): + p0 = self.run1(0.0, select='ep_waste') + p1 = self.run1(0.0, select='endpoint') + torch.testing.assert_close(p0['traj'], p1['traj'], rtol=0, atol=0) + def test_zero_weight_is_the_old_endpoint_path(self): """len_w=0 且 select='ep_len' 必须与 select='endpoint' 逐位相同（不许悄悄改默认行为）""" class SafeKin: